1
$\begingroup$

Let $x_1,\dots,x_n$ be i.i.d. drawn from $N(0,I_{p\times p})$. Consider the sample covariance matrix $W(n,p)=\frac 1n \sum_{i=1}^n x_ix_i^T$, a Wishart matrix.

For fixed $n,p$, what is the expected spectral norm of $W(n,p)$, $$\mathbb E \left[ \left\|\frac 1n \sum_{i=1}^n x_ix_i^T\right\|_2\right ] ?$$

I'm familiar with the asymptotic Bai/Silverstein result, but not sure whether that can be exploited in the finite case. If the exact expression is unwieldy, a reasonable upper bound may suffice.

$\endgroup$
  • 2
    $\begingroup$ Cross-posted to stats.stackexchange: stats.stackexchange.com/questions/239916/… Please do not cross-post, especially not after only an hour. Give it a week or so before giving up and posting elsewhere; better would be to flag for migration; at the very least one should notify users, at each site where you post, of all the other SE sites where you post. $\endgroup$ – Todd Trimble Oct 13 '16 at 13:04
3
$\begingroup$

I don't think there's an exact expression, but the Bai–Yin result does give the right prediction. It's a little easier to state nice-looking results for a $p \times n$ matrix $X$ with independent standard Gaussian entries, so that what you're asking about is $n^{-1} \mathbb{E} \| X \|_2^2$. The nicest result is that $$ \| X \|_2 \le \sqrt{n} + \sqrt{p} + t $$ with probability at least $1 - 2 e^{-t^2/2}$ (see Theorem II.13 of this paper or Theorem 5.35 of this paper), which implies in particular that $$ \mathbb E \left[ \left\|\frac 1n \sum_{i=1}^n x_ix_i^T\right\|_2\right ] \le \left(1 + \sqrt{\frac{p}{n}}\right)^2 + \frac{4}{n} e^{-(\sqrt{n}+\sqrt{p})^2/2} $$ assuming I did the integrals right. I don't recall offhand a similarly precise result for the lower bound, though there may be one out there. Chances are, though, that the upper bound is what you really need.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.